Prediction of cardiorespiratory events in preterm infants

Every year, about 15 million infants are born prematurely, and approximately 1 million child deaths are related to complications of preterm birth. Cardiorespiratory events, defined as intermittent apnoea, hypoxemic or bradycardic episodes, are the most com?mon developmental disorder in preterm infants and are associated with some acute and long-term complications, including retinopathy, multi-organ dysfunction, and neurodevel?opmental impairment. If these cardiorespiratory events can be predicted, preventative interventions could be applied to minimise harm to the infant or even avoid these events. Machine learning techniques have already shown their usefulness in healthcare. This thesis aimed to investigate the utility of machine learning models for the automated prediction of cardiorespiratory events.
In this study, data were collected at Royal Hobart Hospital involving 31 preterm infants and included 3591 hours of electrocardiogram and respiratory motion recordings. We extracted a set of features from the signals available in the current neonatal intensive care unit based on feature analysis and the knowledge of domain experts. Changes in the feature distribution at various time intervals preceding these cardiorespiratory events were analysed to identify potential precursors to the cardiorespiratory events. A variety of machine learning techniques has been tried and their performances were compared, including Artificial Neural Network, Support Vector Machine, Random Forest, and some deep-learning methods.
We found that bradycardia prediction would benefit from incorporating respiratory information. Prediction of hypoxemic events using our trained machine learning models outperformed the current threshold-based alarm system. Our feature analysis demonstrated that apnoea does not occur in isolation but is heralded by cardiorespiratory instability. Our findings suggest that some subtle patterns in the cardiorespiratory signals can provide important information for predicting apnoea. Even though there are still a number of challenges that need to be overcome before these predictive models can be applied to real-life healthcare, the prediction of cardiorespiratory events is promising.